Darkstone: lossless telemetry sink tested; allocation refused, 2026-11-12

Run: scratch/runs/20261112-darkstone-lossless-route-2330/.

A distinct storage blocker prevented allocation: the actual admission measurement was 1,497,677,824 bytes free on /home, below the required strict 1,500,000,000-byte floor. No sandbox was created and no original game input was issued. There is no new Ardyl dungeon, NPC, quest or checkpoint progress. This avoids repeating the 2300 run's capacity stop after provisioning and Town travel.

Inspection covered actual parent control.py, provision-gated.py, recover.py, browser.cjs, capture retrieval, and Transport. The actual provisioning caller invokes Transport.provision directly, whose telemetry reaches control.record; the unused control.provision duplicate uploads writer is not this path. Recovery already gzip-compresses large JSONL/module outputs and hardlinks live PNGs. Measured parent retention is 17,009,796 logical bytes across 342 paths, but only 12,267,520 allocated bytes across 328 unique inodes. PNGs account for 6,152,192 allocated bytes. Duplicate path counts therefore do not represent reclaimable PNG copies.

Only new run-local helpers changed. The copied control.record callback now writes the same raw JSONL bytes as complete concatenated gzip members under a process lock, flushed and fsynced before returning. Original terminal side receipts, Transport identity/recovery/noReplay behavior, ACKs, reviews, capture writers, and parent/EOF receipts are unchanged. Seals retain full decoded size/SHA-256, encoded size/SHA-256 and record count; whole-stream hash binds byte order. A missing seal is unqualified and later appends invalidate it. Truncated gzip or incomplete JSONL is rejected without repair or replay. This is an evidence sink change, not a change to command outcome handling.

Actual emitted producer/consumer tests used the production Transport and the changed control callback against a local HTTP fixture. A successful command emitted 1,513,043 raw bytes, retained as 5,286 compressed bytes, with full exact roundtrip. Separate UNKNOWN, truncated-output and disconnected-status cases preserved terminal failures and rejected a second command before POST. Damaged gzip members and partial JSONL were rejected. Reencoding all 1,770 actual historical parent records preserved 1,160,215 raw bytes as 550,878 encoded bytes, saving 609,337 bytes relative to live raw retention. These are local tests, not new original gameplay or input qualification.

That prospective saving cannot restore current headroom: the parent's 1,076,614-byte provision prefix already occupies only 90,555 gzip bytes. Its remaining raw API continuation is only 83,601 bytes. The new per-event encoding is intentionally less compact than that existing whole closed-prefix gzip; replacing the parent gzip would increase storage. No parent artifacts were rewritten/unlinked, and no cache, reserve, vendor, media, session or other worker outputs were touched. All parent files were hashed in the measurement inventory. No production source changed, so no production build is claimed; helpers are recoverable from owned-inputs.patch plus exact Git state, with decoding instructions in RECOVERY.md.

No new box ID, expiry or stop receipt exists. Historical parent bx_k3aptuf8 remains recorded as archived, expiry null, normal completed unforced stop stop_665faaa9e25f3a9c at 23:05:55.195Z; this task did not requery, resume or delete it. Parent outer lifecycle exit143/missing waits remain UNKNOWN. Existing complete six-file living fallback, original media manifest and exact qualified input helpers remain untouched and authoritative.

Next: obtain actual free space above the floor, use this tested sink for a fresh watched private original, and continue beyond the Town approach toward an actual dungeon or quest NPC using the supported map button. Reuse the exact complete six files and verify production mounting/raw readback and native 35/35 life, food82, gold393 preview. The 2300 map opening is historical evidence only.